Evaluation of different convolutional neural network encoder-decoder architectures for breast mass segmentation

In this work, we study convolutional neural network encoder-decoder architectures with pre-trained encoder weights for breast mass segmentation from digital screening mammograms. To automatically detect breast cancer, one fundamental task to achieve is the segmentation of the potential abnormal regions. Our objective was to find out whether encoder weights trained for breast cancer evaluation in comparison to those learned from natural images can yield a better model initialization, and furthermore improved segmentation results. We applied transfer learning and initialized the encoder, namely ResNet34 and ResNet22, with ImageNet weights and weights learned from breast cancer classification, respectively. A large clinically-realistic Finnish mammography screening dataset was utilized in model training and evaluation. Furthermore, an independent Portuguese INbreast dataset was utilized for further evaluation of the models. 5-fold cross-validation was applied for training. Soft Focal Tversky loss was used to calculate the model training time error. Dice score and Intersection over Union were used in quantifying the degree of similarity between the annotated and automatically produced segmentation masks. The best performing encoder-decoder with ResNet34 encoder tailed with U-Net decoder yielded Dice scores (mean +/- SD) of 0.7677 +/- 0.2134 for the Finnish dataset, and ResNet22 encoder tailed with U-Net decoder 0.8430 +/- 0.1091 for the INbreast dataset. No large differences in segmentation accuracy were found between the encoders initialized with weights pre-trained from breast cancer evaluation, and of those from natural image classification.